Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Who matters in coordination problems on
networks: myopic or farsighted agents?
Ana Mauleon∗ Simon Schopohl† Akylai Taalaibekova‡
Vincent Vannetelbosch§
January 2019
Abstract
This paper studies a model of social interaction in a fixed network where agents
play a coordination game - a game where it is optimal for a player to choose an
action like most of her friends. The different actions correspond to different projects
the player can invest into. A project is successful once a certain amount of players
have chosen it. All players have a certain type: A player can be either an extremist
for one of the projects or she can be a moderate. Extremist players only obtain
utility from one project, while moderate players are ex ante indifferent between the
projects. In addition, the players may also differ in their level of farsightedness:
Some players cannot foresee the reactions that their actions cause while others
anticipate all induced changes.
We propose a solution concept to study the strategy profiles that are stable when
myopic and farsighted players interact with each other. In such profile no player has
any incentive to deviate to another action. We show how the set of stable strategy
profiles changes if a myopic player becomes farsighted, a moderate player into an
extremist or vice versa. Furthermore, we characterize the set of stable strategy
profiles for common network structures.
Keywords : Coordination game, extremism, heterogeneous players, myopic and farsighted
players, stability.
JEL Classification : C70, D83, D85.
∗UCLouvain. E-mail: [email protected]†UCLouvain. E-mail: [email protected]‡UCLouvain and Universite Paris 1. E-mail: [email protected]§UCLouvain. E-mail: [email protected]
1 Introduction
Coordination problems play a huge role in our everyday life. We coordinate with our
partners and family what to eat, how to spend the evening or where to go on vacation.
With co-authors we have to coordinate time and date of meetings and the splitting of tasks.
In economics, coordination games have been studied under different aspects. In this paper
we analyse a coordination game that is played between players who are connected by a
network. The players can choose between different projects and while some players may
be ex ante indifferent between all projects, other players are very strongly in favor of a
certain project. Each player wants to invest into a project she likes, but at the same
time she wants to coordinate with her neighbors in the network. The projects the players
invest to can either be successful or not. For a project to be successful enough players
have to choose the project such that the threshold is reached. Only then the players get
a positive payoff. Those thresholds can differ between projects. We assume that players
can only choose a single project or they can decide to abstain. Players are heterogeneous
in their taste for the different projects and, in addition, heterogeneous in their degree of
farsightedness. A player can either be myopic or completely farsighted. Myopic players
do not anticipate that individual deviations can yield to subsequent deviations while
farsighted players forecast how others might react to their actions.
In this paper we characterize the set of stable strategy profiles. In a stable profile all
players have chosen a project (or abstention) and no player has any incentive to deviate.
A player deviates from a profile if her utility in the resulting profile is higher. A myopic
players only compares the current profile to the profile in which only she deviates, while
a farsighted player anticipates that other players might adjust their actions as well.
We show that the set of stable profiles is the set of Nash equilibria if all players are
myopic. Modifying the degree of farsightedness of a player such that this player becomes
farsighted reduces the size of the set and works as an equilibrium refinement. We prove
that even if all players are farsighted the set of stable profiles is non-empty, which implies
that for all combinations of myopic and farsighted players the set is non-empty.
Furthermore, we analyse the impact of extremism. A player is an extremist for a
project if her taste between the projects is such that she only likes one project. An
extremist will only invest into that project, or abstain. We state results how the set of
stable profiles changes when the taste of a player is modified. In addition, we show how
extremists can divide networks into different parts which can be analysed separately.
Besides the farsightedness and the extremism, the thresholds of the projects influence
the set of stable profiles. We characterize the change in the set of stable profiles if we
increase the threshold of a project. In that case some profiles may be not stable anymore,
but other profiles may become stable.
We find many different examples in the literature and applications in our everyday
life. Example 1: The classic example for coordination games is the decision of a couple
1
about their evening activity. In this battle of the sexes the woman prefers going to the
opera with the man over watching a football match, while the man’s preferences are the
opposite way. He prefers football over going to the opera. No player gets any utility from
spending the evening apart. We can adjust this example to work in our model: Let the
network be a social network where links imply that the linked players are friends. All
players have to make a decision about their evening activity. They can either stay at
home (abstention) or choose one of several projects. They can go to the cinema, play
darts, bowling, football or go for drinks and they enjoy the activity the most if all their
friends are there as well. Clearly it is possible to go to the cinema or for drinks alone,
but for playing darts or bowling you need at least two players and for football even more.
Some players may spend each evening at the bar, they are extremists for that project. A
player that decides to go and play football only and finds that no one else is there receives
a negative utility. She would have preferred to stay at home over wasting her time to go
to the pitch.
Example 2: Another example is the cooperation between researchers (or R&D firms).
Clearly, the researchers build a network, where the links can be interpreted as previous
cooperation. The players can do research into different problems. While some problems
might be solved by a single player, other problems need the cooperation of several players
to be solved. If there are not enough players to solve the problem, they do not get any
result and have a negative utility. In this example the cooperation with the neighbors
might not give higher payoff to the researcher directly, but lowers the effort she has to
use for her research. In established research cooperations, the players know how the
others work and how to explain things. While some researchers are interested in several
problems, other researchers are extremists for a certain problem and only want to solve
this one problem.
Example 3: Our model can be applied to social networks like Twitter as well. The
users can choose between different hashtags for their posts and want to be part of a viral
campaign, i.e. a hashtag that is used by a large group of users. If they use a hashtag
that is not used often their post does not reach many people. If they are part of a viral
campaign, each user benefits from followers that use the same hashtag as they might
retweet the others posts.
Example 4: The choice of different problems can also be seen as the adoption of new
technologies. Cities may choose one of several ways to make public transport more eco-
friendly and benefit from exchange with neighboring cities. If one of these new types of
transport is not supported by enough cities the development might be stopped. In that
case the cities are not able to further improve their public transport.
Similarly the choice of mobile phone operation systems work. Users benefit from friends
that can help them with problems, but if there are not enough users the developing
company might stop further development as it happened with Windows Phone.
Coordination games have been analyzed intensively in the literature. The battle of
sexes already appeared in Luce and Raiffa (1957) and is used as an example also in
2
Fudenberg and Tirole (1991).
Closest to our work, but without the network aspect is the work of Sakovics and Steiner
(2012). They focus on a continuum of players where each player can either decide to invest
into a project or to abstain from investment. The project has an investment threshold
and is only successful if enough players invest. The players receive a noisy signal about
this threshold. The authors show that depending on the noise in the signal there can be
miscoordination, while if the noise is small the players coordinate their actions.
Jackson and Watts (2002a) model coordination games on networks, but focus on re-
peated interaction. In each period a player can adjust her action. With a small probability
a player makes a mistake which might lead to changes of other players. All players are
myopic and have the same taste between the different options. The authors use a different
stability concept (stochastically stable) and in addition to analyzing the game, Jackson
and Watts study the network formation when links are costly.
Goyal and Vega-Redondo (2005) study a coordination game which is repeatedly played
on networks where players can choose between two different options. Players can revise
their strategy in each period with a certain probability. In their model the links between
players are paid only by one player. The authors show that depending on the costs for a
link the players either coordinate on the risk-dominant action or on the efficient action.
Leister et al. (2018) analyze a global game played on a network. The players receive
a noisy signal about the value of a technology and can choose whether to adopt the
technology or not. Similar to our model, the value is increasing in the amount of neighbors
who adopt the technology as well. The authors show under which conditions all players
(or groups of players) coordinate their actions.
In our examples players have incentives to coordinate with their neighbors, but in other
applications the opposite might be the case. The model of Bramoulle (2007) focuses on
these anti-coordination games. The author defines frustration as a function of neighbors
who play the same action and neighbors who choose the different action. This frustration
is the highest in complete networks and lowest in bipartite networks. Bramoulle et al.
(2004) analyze a similar model with network formation.
More specific models with network effects similar to coordination games are for exam-
ple for public goods Allouch (2015) and Bramoulle and Kranton (2007) or the model of
spill-overs by Ballester et al. (2006).
In our model we will often observe communities that choose the same action. Recent
work on communities in networks has been done by Jackson and Storms (2018). They
study a game in which a player can either adopt the behavior of her neighbors or not. If
a certain amount of neighbors of a player choose to adopt, it is beneficial for the player
to adopt as well. Several equilibria might exist and the authors define a community as a
set of players that coordinates in each equilibrium.
Our model uses the idea of improving paths as in Jackson and Watts (2002b). Far-
sightedness has been studied in many different models and applications, see for example
Ray and Vohra (2015) or Herings et al. (2009).
3
The structure of our paper is as follows: In Section 2 we describe the model. Section
3 defines the stability concept and characterises the set of stable strategy profiles when
all players are myopic. In Section 4 we analyze the impact of farsightedness, extremism
and the thresholds on the strategy profiles. We prove in Section 5 that this set of stable
profiles is never empty. Results for specific networks are given in Section 6. In Section
7 we discuss the incentives and possibilities of a social planner. Section 8 concludes. All
proofs are relegated to the appendix.
2 Model
Let N = {1, ..., n} be a finite set of players, which are arranged in an undirected network
g. A link between two players i and j is denoted by ij ∈ g. By Ni(g) = {j ∈ N : ij ∈ g}
we denote the set of neighbors of player i in the network g. A player who only has one
neighbor is called a loose end. The complete network is the network in which each player
i is linked to all other players, i.e. Ni(g) = N \ i. In the star network there is one
player i in the center of the star and all other players j are only linked to the center, i.e.
Ni(g) = N \ i and Nj(g) = {i}. The line network is the network in which each player
i ∈ {2, . . . n−1} is linked to the players i−1 and i+1 and players 1 and n are loose-ends.
Player 1 is only linked to player 2 and players n only link is to player n− 1.
We study a coordination game that is played by all players. There exist r different
projects. The set of projects is denoted by P = {P 1, . . . , P r}. A project Q ∈ P is suc-
cessful if at least EQ players choose this project. We call EQ ∈ {1, . . . , n} the investment
threshold of project Q. Let E = (E1, ..., Er) denote the vector of investment thresholds.
All players in the network choose a project or abstain. The choice of player i is captured in
pi, i.e. pi ∈ P = P ∪ ∅ = {∅, P 1, . . . , P r}, where pi = ∅ means that player i abstains. We
denote a strategy profile by the vector1 p = (p1, . . . , pn). The set of all possible strategy
profiles is P = P n = {∅, P 1, . . . , P r}n.
Players have a taste for the different projects, where tQi ∈ {0, 1/r, 1} denotes the taste
of player i for project Q with∑r
Q=1tQi = 1. We assume that a player can either be an
extremist for one project Q, i.e. tQi = 1 or be moderate, i.e. tQi = 1/r for all Q ∈ P .
The set of extremists for project Q is denoted by XQ = {i ∈ N : tQi = 1} and the set of
moderate players by Y = {i ∈ N : tQi = 1/r ∀Q ∈ P} . Let t = ((tQ1 )rQ=1, ..., (t
Qn )
rQ=1)
denote the tastes of the players for the different projects.
In addition to their different tastes, players can vary in their degree of farsightedness.
Let si = {m, f} denote the degree of farsightedness of player i. If si = m, player i is
myopic, while player i is farsighted if si = f . We denote by M the set of myopic players
and by F the set of farsighted players. Obviously, we have M ∪F = N . Either set (M or
F ) is allowed to be empty. The taste of player i for the different projects (tQi )rQ=1 and her
1For a n × 1 vector a we denote by a−i the vector of size n − 1 with the i-th entry removed and
analogue we use a−i,j to denote the vector in which the i-th and the j-th elements have been deleted.
4
degree of farsightedness si determines the type of player i, θi = ((tQ1 )rQ=1
, si). We denote
by θ = (θ1, ..., θn) the vector of types of the different players.
The utility function ui of player i, given (g, θ, E), is assumed to be as follows:
ui(p) =
0 if pi = ∅, (1a)
< 0 if pi 6= ∅ and either |{j ∈ N | pi = pj}| < Epi or tpii = 0, (1b)
> 0 if pi 6= ∅ and |{j ∈ N | pi = pj}| ≥ Epi and tpii > 0. (1c)
Properties of u:
(P1) ui(pi, p−i) = ui(pi, p′−i) for any p−i and p′−i if ui(pi, p−i) and ui(pi, p
′−i) as in (1a) or
(1b).
(P2) ui(pi, pj, p−i,j) > ui(pi, p′j, p−i,j) if p
′j 6= pj = pi and ui as in (1c).
(P3) ui(pi, pj, p−i,j) = ui(pi, p′j, p−i,j) if pj 6= pi 6= p′j.
(P4) ui is independent of si and s−i.
We normalize the utility of abstention to 0 (1a). An investment of player i into an
unsuccessful project gives the player a negative utility (1b), while an investment into a
successful project Q yields the player a positive utility as long as the taste of player i
for this project Q is positive (1c). Furthermore, in that case, the utility of player i is
increasing in the amount of neighbors in the network g who choose the same project (P2).
In the other cases (1a and 1b) the utility of player i does not change if her neighbor,
player j changes her action as long as the utility of player i stays non-positive (P1).
Similarly, the utility of player i who invests in a successful project pi does not change if
her neighbor j invests in project pj 6= pi or p′j 6= pi (P3). We can also see that the utility of
a player depends on her taste and project choice, but not on her degree of farsightedness.
Furthermore, it follows directly from the utility function that an extremist for project Q
will never choose project R 6= Q.
Example 1. One example for this class of utility function is given by:
ui(p) =
0 if pi = ∅, (2a)
−ci if |{j ∈ N : pi = pj}| < Epi , (2b)
tpii
[
αi + (1− αi)|{j∈Ni:pi=pj}|
|Ni|
]
− ci if |{j ∈ N : pi = pj}| ≥ Epi (2c)
Players have different interests in cooperating with their neighbors. This parameter of
cooperation is captured by αi ∈ (0, 1). The investment into a project is costly for each
player and the cost for player i is denoted by ci. In case of a successful investment the
return to the player is always higher than her investment cost, i.e. αi > r · ci holds.
(2c) corresponds to the case of a successful project pi. In that case the player’s utility
depends on her taste (tpii ) for the project pi, her parameter for cooperation αi and the
amount of neighbors who also invested into project pi. From this payoff the cost of
investment ci has to be subtracted. In case of an unsuccessful project pi (see 2b), there is
no benefit from the project, but only the cost ci. (2a) represents the case of abstention.
5
From now on we only consider projects that can be successful, i.e. ∀Q ∈ P : |{i ∈
N : tQi > 0}| ≥ EQ. This ensures that a project Q can be successful and that there are
not too many extremists for all the other projects such that the investment threshold of
project Q cannot be reached even if all extremists for project Q and all moderate players
invest into project Q.
❍❍❍✟✟✟
✟✟✟❍❍❍
❅❅❅
A
A
✉t✉t
✉t
✉
B
1
2
3
4
5
6
7
Figure 1: Network example
In Figure 1 we illustrate the notation we use for figures. Players with a round node
are moderate players, while a player with a square node is an extremist for the project
written in the square. If the node is filled the player is farsighted otherwise she is myopic.
So in this example there are two extremists for project A, namely player 1 and 2. Player
7 is an extremist for project B and the players 1 and 6 are farsighted.
3 Myopic-farsighted stability
For normal form games, the stability concept studied in the literature is the concept of
Nash equilibrium.
Definition 1. Consider (g, θ, E) as given. A strategy profile p∗ ∈ P is a Nash Equilibrium
if it satisfies
∀i ∈ N, ∀pi ∈ P : ui(p∗i , p
∗−i) ≥ ui(pi, p
∗−i)
Let N (g, θ, E) denote the set of all Nash Equilibria.
The notion of Nash equilibrium considers that players are myopic in the sense that
they do not forecast how others might react to their actions. This concept is based on
the direct dominance relation and neglects the destabilizing effect of indirect dominance
relations as introduced by Harsanyi (1974) and Chwe (1994). Indirect dominance captures
the idea that farsighted players can anticipate the actions of other players and consider
the end strategy profile that their deviations may lead to. Based on the concept of indirect
dominance, Chwe (1994) has proposed the largest consistent set to predict which of the
several pure strategy Nash equilibria will be played if players are farsighted. Up to now,
no solution concept has been proposed in order to allow for heterogeneity in the degree
of farsightedness among players. In the following, we propose a solution concept to study
the strategy profiles that are stable when myopic and farsighted players interact with each
other: the set of myopic-farsighted stable strategy profiles.
A myopic-farsighted improving path is a sequence of strategy profiles that can emerge
when farsighted players choose an action based on the improvement the end strategy
6
profile offers them relative to the current strategy profile while myopic players choose an
action based on the improvement the next strategy profile in the sequence offers them
relative to the current one. Each strategy profile in the sequence differs from the previous
one in that only one action of one of the players has been modified.
Definition 2. Consider (g, θ, E) as given. A myopic-farsighted improving path of length
L from a strategy profile p ∈ P to a strategy profile p′ ∈ P is a finite sequence of strategy
profiles p0, . . . , pL ∈ P with p0 = p, pL = p′ and pj 6= pk for all j, k ∈ {1, . . . , L − 1}
such that for every ℓ ∈ {0, . . . , L− 1} there is a unique player i such that pℓ+1
i 6= pℓi and
ui(pℓ+1) > ui(p
ℓ), if i ∈ M
ui(pL) > ui(p
ℓ), if i ∈ F.
If there exists a myopic-farsighted improving path from the strategy profile p to p′ we
write p → p′. The set of all strategy profiles that can be reached from p by a myopic-
farsighted improving path is denoted by h(p) = {p′ ∈ P | p → p′}.
Example 2.
❍❍❍✟✟✟
✟✟✟❍❍❍
❅❅❅
A
✉t✉t
✉t
✉t
✉t
B
1
2
3
4
5
6
7
Figure 2: Network example
For this example we consider the network shown in Figure 2. All players are myopic.
Players 1 and 7 are extremists while all other players are moderate. There are only two
projects A and B and both have low thresholds, i.e. EA = EB = 1. Consider the profile
p0 = {A,B,B,A,A,A,B}. The players 1, 4, 5 and 6 chose project A and the other
players selected B. One example for an improving path out of this profile is the following:
Player 3 has only one neighbor (player 2) with the same choice as her and three neigh-
bors who chose project A. So it is an improvement for her to select A as well. From this
profile (p1 = {A,B,A,A,A,A,B}) there is the second step to p2 = {A,A,A,A,A,A,B}.
Caused by the change of player 3, player 2 is not coordinating with any neighbor. She
improves if she changes her choice to project A.
In this example all players are myopic and only consider the immediate improvement
they get from changing their action. This is different if we consider farsighted players.
Let us take once again the example from Figure 2, but assume that player 6 is farsighted.
In the profile p0 = {A,A,A,A,A,A,B} player 6 is coordinating with one of her two
neighbors. From this profile there is an improving path initiated by player 6. First, player
6 will change to project B, which yields to the profile p1 = {A,A,A,A,A,B,B}. In this
profile player 4 has an incentive to deviate to project B as well, because two of her three
neighbors have selected project B. This yields to the profile p2 = {A,A,A,B,A,B,B}.
7
We see that in p2 player 6 is coordinating with both of her neighbors. Since she is
farsighted, she had the incentive to start the improving path, taking into account that
player 4 will follow her example and deviate from project A to B. This makes player 6 in
profile p2 better off than in p0, even though she did not immediately (in p1) benefit from
her change.
Only farsighted players anticipate the changes of the other players and compare their
current utility with the utility they would get at the end of the improving path.
Nash equilibrium requires strategy profiles to be immune to immediate deviations
and does not capture that farsighted players anticipate the actions of other players and
consider the end strategy profile that their deviations may lead to. In order to be stable,
a strategy profile should be immune to deviations of both myopic and farsighted players.
These considerations lead to the following definition of myopic-farsighted stable strategy
profiles.
Definition 3. Consider (g, θ, E) as given. A strategy profile p is myopic-farsighted stable
if it satisfies h(p) = ∅. The set of myopic-farsighted stable strategy profiles Z(g, θ, E) is
defined as follows: Z(g, θ, E) = {p ∈ P : h(p) = ∅}.
According to Definition 3, a strategy profile p is myopic-farsighted stable if there is no
myopic-farsighted improving path leaving p. Let us first analyze some properties of the
set of myopic-farsighted stable strategy profiles.
The next proposition establishes that there cannot be a myopic-farsighted improving
path that starts at a strategy profile p and ends at the same strategy profile p. The result
holds when both myopic and farsighted players deviate along the path.
Proposition 1. For any strategy profile p ∈ P, it holds that p /∈ h(p).
Notice that if p∗ ∈ N (g, θ, E) is a Nash equilibrium, then there is no profitable indi-
vidual deviation from p∗. Thus, h(p∗) = ∅, when all players are myopic. As a corollary of
Proposition 1, we have that there always exist a myopic-improving path from any strategy
profile p /∈ N (g, θ, E) to some Nash equilibrium strategy profile p∗ ∈ N (g, θ, E).
Corollary 1. Assume all players are myopic. Then, for any strategy profile p /∈ N (g, θ, E),
there exists a Nash equilibrium strategy profile p∗ ∈ N (g, θ, E) such that p∗ ∈ h(p).
From Proposition 1 and Corollary 1, we obtain the next proposition.
Proposition 2. If all players are myopic, a set of strategy profiles is a set of myopic-
farsighted stable strategy profiles if and only if it is equal to the set of Nash equilibrium
strategy profiles.
Proposition 2 states an important result as it shows the connection between the set
of myopic-farsighted stable strategy profiles and the set of Nash equilibria. As long as all
players are myopic the sets are the same.
8
4 Farsightedness, Extremism and Thresholds
In the previous section we have characterized the set of stable strategy profiles for the
case that all players are myopic. Besides the network, the myopic-farsighted stable set
Z(g, θ, E) depends on three different sets of parameters: It plays an important role
whether a player is myopic or farsighted and if she is moderate or an extremist. In
addition, the thresholds of the projects have a big impact on the myopic-farsighted sta-
ble set. In this section we characterize the impact of farsightedness, extremism and the
thresholds and answer our question “Who matters in coordination games?“.
We first state some remarks that simplify our analysis.
Remarks 1.
• In all stable profiles with at least one successful project, all moderate players will
choose a successful project and will never abstain.
• Assume player i is an extremist for project Q. In all stable strategy profiles where
project Q is successful, player i will choose project Q.
• Assume player i is moderate and a loose-end and player j is the neighbor of player
i. In all stable strategy profiles where j does not abstain, player i will choose the
same project as player j.
• If all thresholds are low, i.e. EQ = 1 for all Q ∈ P no player will abstain in a stable
profile.
• If the thresholds are such that only one project can be successful (i.e. EQ+ER > n
for all Q,R ∈ P ), then there exist only two different types of stable profiles: Either
one project Q is successful and all extremists for Q and all moderate players choose
Q, while the other extremists abstain or all players abstain.
4.1 Impact of Farsightedness
What happens with the set of myopic-farsighted stable strategy profiles once some play-
ers become farsighted? Our next proposition shows that Z(g, θ, E) weakly decreases
(increase) when a myopic (farsighted) player i becomes farsighted (myopic) while keeping
the types of the other players (θ−i) constant.
Let Z(g, ((tQi )rQ=1
, m), θ−i, E) be the set of stable myopic-farsighted strategy profiles
when player i is myopic and let Z(g, ((tQi )rQ=1, f), θ−i, E) be the set of stable myopic-
farsighted strategy profiles when player i is farsighted. The type of the rest of the players
does not change.
Proposition 3.
Z(g, ((tQi )rQ=1, f), θ−i, E) ⊆ Z(g, ((tQi )
rQ=1, m), θ−i, E).
Proposition 3 implies that the set of stable profiles weakly decreases when a myopic
player turns farsighted. In Section 5 we show that even though the set weakly decreases it
will never become empty. One immediate conclusion from Proposition 3 is that, if players
9
i and j become farsighted, the order in which they become farsighted does not matter for
the set of stable strategy profiles as illustrated in Figure 3.
Z(g, ((tQi )rQ=1, m), ((tQj )
rQ=1, m), θ−i,j, E)
⊆ ⊆
Z(g, ((tQi )rQ=1, f), ((t
Qj )
rQ=1, m), θ−i,j, E) Z(g, ((tQi )
rQ=1, m), ((tQj )
rQ=1, f), θ−i,j, E)
⊆ ⊆Z(g, ((tQi )
rQ=1, f), ((t
Qj )
rQ=1, f), θ−i,j, E)
Figure 3: Impact of farsightedness
From Proposition 3 we obtain the following corollary.
Corollary 2. The set of myopic-farsighted stable strategy profiles when there are both
myopic and farsighted players Z(g, θ, E) is a subset of the set of Nash equilibrium strategy
profiles N (g, θ, E): Z(g, θ, E) ⊆ N (g, θ, E).
Thus, in the type of coordination games studied in the paper, the set of myopic-
farsighted stable strategy profiles when players are both myopic and farsighted is a refine-
ment of the set of Nash equilibrium strategy profiles.
Example 3.
❍❍❍✟✟✟
✟✟✟❍❍❍
❅❅❅
A
✉t✉t
✉t
✉t
✉t
B
1
2
3
4
5
6
7
Figure 4: Network example
All players are myopic, two project (A and B) with
EA = EB = 1. Stable strategy profiles:
1. {A,A,A,A,A,A,B}, 2. {A,A,A,A,B,A,B},
3. {A,A,A,B,A,B,B}, 4. {A,A,A,B,B,B,B},
5. {A,A,B,B,B,B,B}, 6. {A,B,B,B,B,B,B}.
Let us consider the network we have already used in previous examples and which is
shown in Figure 4. We consider only two project A and B and thresholds EA = EB = 1.
If all players are myopic the profiles shown on the right are stable. Hence, the set of Nash
equilibria contains exactly those six profiles.
Turning any of the players 1 to 5 into farsighted players, does not change the stable
set at all. On the other hand, if player 6 becomes farsighted, the first two profiles are not
stable any more. In those profiles player 6 deviates from project A to project B and takes
into account that player 4 then will change to B as well (as explained in Example 2).
Even when all players are farsighted the profiles 3 to 6 remain stable.
4.2 Impact of Extremism
In the previous results we have assumed that the taste of the players is fixed. In the
examples, there was one extremist for project A, one extremist for project B and the
10
remaining players were moderate. We can clearly observe that the set of myopic-farsighted
stable strategy profiles does not only depend on the farsightedness of the players, but also
on their taste.
Let Z(g, ((tQi = 1/r)rQ=1, si), θ−i, E) be the set of myopic-farsighted stable strategy
profiles when player i is moderate, and Z(g, ((tQi = 1, (tRi = 0)R6=Q), si), θ−i, E) be the set
of myopic-farsighted stable strategy profiles when player i is an extremist for project Q.
Proposition 4.
Z(g, ((tQi = 1/r)rQ=1, si), θ−i, E) ⊆⋃m
Q=1Z(g, ((tQi = 1, (tRi = 0)R6=Q), si), θ−i, E)
Proposition 4 can be interpreted as follows. If choosing a project Q is part of a stable
strategy profile p when player i is moderate, then the strategy profile p is still stable once
player i becomes an extremist for project Q, while keeping constant the type of the other
players. If there exists no stable strategy profile in which player i, while being moderate,
choose project Q, profiles become stable when player i is an extremist for project Q.
Example 4. Let us consider the same network structure as in the previous examples and
let us assume that all players are myopic and thresholds are low, i.e. EA = EB = 1.
In the case that all players are moderates, there are 8 different stable strategy profiles:
p1A = {A,A,A,B,A,B,B}, p2A = {A,A,A,B,B,B,B}, p3A = {A,A,B,B,B,B,B}, p4A =
{A,A,A,A,A,A,A} and p1B, . . . , p4B which are the same profiles with the actions A and
B exchanged.
Turning player 7 into an extremist for project B keeps only those profiles stable in
which player 7 chooses project B, namely: p1A, p2A, p
3A and p4B. In addition there are
new stable strategy profiles: p5 = {A,A,A,A,A,A,B}, p6 = {A,A,A,A,B,A,B} and
p7 = {B,B,A,A,A,A,B}. We see that the profiles p5 and p7 are modifications from the
profiles p4A respectively p3B: The actions of the players 1 to 6 in profile p5 (p7) are the
same as in p4A (p3B) only the action of player 7 is different. In the strategy profile p6 there
is a deviation from at least two players compared to the profiles p1A to p4A and p1B to p4B.
If, in addition, we make player 1 an A-extremist we get the same six stable strategy
profiles as in Example 3: Profiles p1A, p2A, p
3A, p
5, p6 and p8 = {A,B,B,B,B,B,B}. Again,
turning a player into an extremist makes some strategy profiles unstable, while it stabilizes
others.
On the other hand turning a player into an extremist does not always create new
stable strategy profiles:
Example 5. We use again the same network and thresholds as in the previous examples.
As we have seen if all players are myopic and players 2 to 6 are moderate, while player
1 (7) is an extremist for A (B), there exist six stable strategy profiles (see Example 3).
If we turn player 3 into an A-extremist only those four profiles where she choose project
A remain stable. The same happens when we turn her into an extremist for project B:
There are only two stable strategy profiles: Those in which player 3 invests into project
B.
11
Proposition 5.
Assume the thresholds of all projects are low, i.e. EQ = 1 for all Q ∈ P . If on all paths
between two players i and j there is an extremist player, the choice of player i and j are
independent of each other.
If we are analyzing the game when the thresholds are not important we can simplify
our analysis with the help of Proposition 5. If there are two (or more) groups of players
who are only connected to the other group(s) through paths that lead through at least
one extremist we can split the graph into two (or more) components and analyze them
separately. One example is illustrated in Figure 5. Players 1 and 6 are only connected
through extremists (players 4 and 5), so instead of analyzing the stable profiles in g we
can separately analyze g1 and g6.
✟✟✟
❍❍❍
❅❅❅
❍❍❍
✟✟✟✈✉
✈✉
✈✉ A
B
✈✉1
4
5
6
(a) g
✟✟✟
❍❍❍
❅❅❅
✈✉
✈✉
✈✉ A
B1
4
5
(b) g1
❍❍❍
✟✟✟A
B
✈✉
4
5
6
(c) g6
Figure 5: Illustration of Proposition 5.
4.3 Impact of the Thresholds
In addition to the characteristics of the players, the thresholds of the project have an
impact on the stability of profiles. If we increase the threshold of project Q from EQ to
EQ + 1 certain strategy profiles will not be stable, but other profiles can become stable.
In profiles in which a project’s threshold was exactly reached the increase in the threshold
makes the project unsuccessful and the profile unstable.
Proposition 6.
Z(g, θ, (EQ + 1, (ER)R6=Q))
⊆ Z(g, θ, E)\{p ∈ Z(g, θ, E) : |{i ∈ N : pi = Q}| = EQ}
∪ {p : pi = ∅ if tQi = 1 and pi ∈ P\Q otherwise}
Proposition 6 shows that an increase of the threshold of project Q from EQ to EQ+1
has the following effect: The new set of stable profiles consists of all those profiles that
were stable before and where at least EQ + 1 players chose project Q. In addition, the
set might contain profiles in which no player chooses project Q and where the extremists
for project Q abstain.
All profiles in {p ∈ Z(g, θ, E) : |{i ∈ N : pi = Q}| = EQ} can no longer be stable if
the threshold is EQ+1. In those profiles the threshold for project Q was reached exactly,
so an increase in the threshold makes project Q unsuccessful. In that case it is better for
all players who invested into Q to choose another project or to abstain. These profiles
12
are described in the set {p : pi = ∅ if tQi = 1 and pi ∈ P\Q otherwise}. Notice that there
might be some strategy profiles that are part of the first and second set.
Example 6. We use, once more, the network from Example 3 with only myopic players.
In the case of low thresholds (EA = EB = 1) there are the six stable profiles as stated
before:
p1 = {A,A,A,A,A,A,B}, p2 = {A,A,A,A,B,A,B}, p3 = {A,A,A,B,A,B,B},
p4 = {A,A,A,B,B,B,B}, p5 = {A,A,B,B,B,B,B}, and p6 = {A,B,B,B,B,B,B}.
An increase of the threshold for project A to EA = 3 (while keeping EB = 1) makes
all profiles unstable in which one or two players have selected project A. Only the profiles
p1 to p4 remain stable. In p5 player 2 will deviate to project B, while player 1 abstains.
Player 1 will also change to abstention in profile p6. This creates one new stable profile:
p7 = {∅, B, B,B,B,B,B}.
An additional increase of EB to EB = 2, also makes profile p1 unstable. Instead
the profile p8 = {A,A,A,A,A,A, ∅} becomes stable. In addition there is another stable
profile: The profile in which all players abstain: p9 = {∅, ∅, ∅, ∅, ∅, ∅, ∅}. This profile is
stable, because all players are myopic. In this profile no player will deviate to a project,
because the investment of a single player is not enough to make the project successful.
The previous example has shown that the profile in which all players abstain can be a
stable strategy profile if the thresholds are high and there are no farsighted players. The
following Proposition characterizes exactly when this profile is stable.
Proposition 7.
The strategy profile in which all players choose to abstain is stable if and only if for all
project Q ∈ P it holds:
|{i ∈ F : tQi > 0}|+min(
1, |{i ∈ M : tQi > 0}|)
< EQ
Proposition 7 gives a certain level of farsighted players that is required to ensure that
the profile in which all players abstain is not stable. There is no improving path out
of the profile in which all players abstain if there are not enough farsighted players in
comparison to the thresholds. To have an improving path, there has to be at least one
project Q for which there exist at least EQ − 1 farsighted players that are not extremist
for a different project. In addition, there either has to be one more player i (farsighted
or myopic) that is not an extremist for a different project. Then, the improving path is
as follows: First the EQ − 1 farsighted players change their decision and choose project
Q. Then, player i choose project Q as well and her decision makes the project successful.
All these players receive a positive utility.
5 Existence of myopic-farsighted stable profiles
We have shown that the set of myopic-farsighted stable strategy profiles Z is the set of
Nash equilibria if all players are myopic. Furthermore, we know from Proposition 3 that
13
the cardinality of the set (weakly) decreases if we replace a myopic player by a farsighted
player. This section shows that Z is never empty and we provide a method to find a
myopic-farsighted stable strategy profile if all players are farsighted.
Starting from an arbitrary profile p, we can start writing the decision-making process
of a player along an improving path as a tree. At each node one player decides whether
she wants to stay with her current decision or wants to deviate to another action. As
players can change their decision many times along the improving path, the decision order
of the players can be neglected. The order only decides which profile we reach, but if there
is an improving path from p when players decide in the order 1 to n, then there is also
one for every other decision order. This is caused by the fact that players can repeatedly
change their decisions. Still, it is not possible to illustrate an entire improving path as a
tree, because there is a infinite number of nodes. Again, the reason is the possibility of
repeated deviations from each player.
The following lemma allows us to write down the decision-making process as a finite
tree if all players are farsighted.
Lemma 1. Assume all players are farsighted. If there exists an improving path from
profile p to profile p′ then there also exists an improving path from p to p′ in which each
player changes her action at most once.
Lemma 1 implies that instead of considering all improving paths, we can focus our
attention on improving paths in which each player changes her action once or not at all.
To find a myopic-farsighted stable strategy profile when all players are farsighted we state
a way to formulate a game tree.
First, we need to fix a decision order, wlog. (1, . . . , n) and we select an arbitrary profile
p0. Then, we can start to create the decision tree. In the first node player 1 decides if she
wants to stay with her action p01 or if she wants to deviate to any other action. In case
of indifference we assume she stays. If she deviates, we arrive in a new profile p1 with
p01 6= p11 and p0−1 = p1−1. In that case we remove player 1 from the decision order and we
start from the beginning of the updated order, in this case with player 2. Player 2 then
faces the same choice, staying in p1 or deviating. If she deviates we move to p2, remove
player 2 from the decision order and start from the beginning of the order. If player i
does not deviate, but stays in profile px, we move to the next player in the decision order
and letting her decide whether she wants to deviate from px. When player i does not
change her decision, she stay in the decision order and can revisit her choice once another
player has deviated. We reach a stable strategy profile either when we reach the end of
the decision order and no player wanted to deviate or when there are no players left in the
decision order. To illustrate this game tree, we introduce the following notation. Let o be
the decision order of the players, e.g. (1, . . . , n) and o \ i implies that player i is removed
from the decision order, i.e. o \ i = (1, . . . , i− 1, i+ 1, . . . n). We generate the game tree
for Γ(p0, o), i.e. starting with profile p0 and order o as shown in Figure 6. Γ(pj, ∅) gives
the utility ui(pj) to player i.
14
③① ③① ③① ③①1
deviate
stay2
s
d
3s s
d
ns
d
u(p0)
Γ(p1, o \ 1) Γ(p2, o \ 2) Γ(p3, o \ 3) Γ(pn, o \ n)
Figure 6: Game tree of Γ(p0, o)
By construction, the finite game tree of Γ(p0, o) shows all the possible improving paths
from p0 in which each farsighted player deviates at most once. Together with Lemma 1
this yields the following result.
Corollary 3. The subgame perfect Nash equilibrium of the game Γ(p0, o) as defined above
is a myopic-farsighted stable profile.
Corollary 3 implies that even if all players are farsighted there always exists at least
one stable strategy profile. Starting from different profiles p0 or using different orders o
can result in different stable strategy profiles, but the main point of this result is to show
that the set of myopic-farsighted stable profiles is non-empty independent of the players’
types.
Example 7. Let us consider the network used in all previous examples and assume all
players are farsighted. We can start from the profile in which all players choose project
A, i.e. po = {A,A,A,A,A,A,A}. We use the decision order following the number of the
players (o = (1, . . . , 7)) and take thresholds EA = EB = 1. Starting from profile p0 the
players 1 to 6 have no incentive to deviate, but player 7 does. She is an extremist for
project B and a profile in which she choose A can never be stable. She changes to project
B and we enter game Γ({A,A,A,A,A,A,B}, (1, . . . , 6)). From this profile the players 1
to 5 again have no incentive to change their projects, but player 6 will deviate. She will
select project B as well and even though this does not yield to an immediate improvement,
she changes, because she takes into account that afterwards player 4 will also deviate to
project B (in the game Γ({A,A,A,A,A,B,B}, (1, . . . , 5)). Those two changes bring us
to the game Γ({A,A,A,B,A,B,B}, (1, 2, 3, 5)). Neither of the players 1, 2, 3 or 5 have
any incentive to change, so the profile is stable.
If we start from the profile {B,B,B,B,B,B,B} only player 1 will deviate and the
resulting profile is {A,B,B,B,B,B,B}, which is stable.
6 Results for specific networks
The stability of a strategy profile does not only depend on the thresholds and the type
of the players, but also on the network structure. In this section we analyze the set of
stable strategy profiles for the most common network structures: The complete network,
the star network and line networks.
15
Complete network In the case of the complete network, each player wants to coor-
dinate with all other players. A moderate player wants to choose the project that gets
chosen by the majority of the other players. This simplifies the analysis.
Proposition 8.
In all stable strategy profiles all moderate players choose the same project, i.e. for all
i, j ∈ Y : pi = pj.
Proposition 8 confirms that the complete network structure leads to a complete coor-
dination of the moderate players. There might be several stable profiles, but in all of them
the moderate players choose the same action. It depends on the thresholds, the amount of
extremists and farsighted players which profiles are contained in the set of stable profiles.
Star network In a star network the center of the star plays an important role. Depend-
ing on the farsightedness and the taste of the player in the center, there can be different
stable strategy profiles. Let player 1 be the center of the star and assume that project B
is the project with most extremists. The following tables show the stable strategy profiles.
t1 moderate extremist for A (wlog)
s1 farsighted myopic myopic
XB > n−1
2XB ≤ n−1
2
EB low EB high and EB low or EB high
p1 B R B R A ∅
pi, i ∈ Y B R B R A R 6= A or ∅
pi, tQi = 1 Q or ∅ Q or ∅ Q or ∅ Q or ∅ Q or ∅ Q or ∅
In addition to the profiles shown, in all cases the profile which all players abstain can
be stable if the conditions from Proposition 7 holds. If there is a moderate, farsighted
player in the center the choice of this player depends on the thresholds of the projects. She
will select the project with the highest number of extremists that also has a sufficiently
low threshold. If the project with most extremists (wlog project B) has a low threshold,
she chooses this project and anticipates that all moderate players and all extremists for
B will do the same. Through this choice she can coordinate with the highest possible
amount of neighbors. If the threshold EB is too high and project B is not successful
player 1 chooses a different project R and the moderate periphery players do the same.
A myopic and moderate center behaves similarly under one condition: More than half
of the periphery players have to be extremists for project B and the threshold of this
project has to be sufficiently low. Then it is the best outcome for the center to choose
project B. If at most half of the periphery players are extremists for B or if the threshold
of project B is too high other stable strategy profiles may exist. In those the player in
the center choose one project R and all other moderate players do the same. The myopic
player in the center is not able to anticipate that the moderate players will always follow
16
her choice. If the center has selected project A and more of the periphery players chose
A than B, the center has no incentive to deviate from her decision.
If the center is an extremist for project A (wlog), there can be only two possible types
of myopic-farsighted stable strategy profiles. Either the center and all moderate players
choose project A (while extremists for different projects abstain or invest into the project
they are extremists for), or the center (and all other extremists for project A) abstains,
all moderate players choose a project R ∈ P\A and all extremists for project Q 6= A
either abstain or choose project Q. In the case where the center abstains, no periphery
player can achieve any coordination. This implies that each periphery player i chooses a
project R which she likes (i.e. tRi > 0) and that is successful. If no project is successful,
all players abstain.
Line networks In the line network the positions of players can be sorted into two cat-
egories: Loose-ends (players 1 and n) and not loose-ends, i.e. players with two neighbors.
We already know that in stable profiles moderate loose-ends will choose the same project
as their neighbor as long as the neighbor does not abstain (Remark 1). On the other
hand, the in-between players can either coordinate with one or two of their neighbors or
with none.
When all players are myopic there exist many different stable profiles. In case of low
thresholds, i.e. EQ = 1 for all Q ∈ P , the line can be fragmented into groups of size 2 or
larger. Players in each fragment choose the same project. If we consider a fragment of
two moderate and myopic players, just turning one of them farsighted will increase the
coordination. The farsighted player will choose the same project as her other neighbor
(from a different fragment) and the myopic player will follow. In a fragment of 3 players we
need to turn at least two of them farsighted in order to make the players in the fragment
coordinate with players from a neighboring fragment. Similarly, if the fragment consist
of x players, at least x − 1 players have to become farsighted to achieve coordination
with the neighboring fragments. The following Proposition provides conditions for full
coordination.
Proposition 9.
Consider a line network with only moderate players. I f there is at most one myopic player
in {2, . . . , n−1} then, there are only r stable strategy profiles ( in which all players choose
the same action). The opposite direction of the result (”only if”) holds if the thresholds
for all projects are low enough, e.g. EQ = 1 for all Q ∈ P .
7 Social Planner
We have seen in the previous sections that the stable strategy profiles can vary in the
amount of successful projects, in the amount of abstaining players and in the coordination
17
between neighbors. From a social planner’s point of view there might be interest to let
the players coordinate as much as possible.
This section discusses two different types of social planner, shows how he can influence
the players choices and yields to the result that the social planner cannot always use the
same receipt for success.
We return to our motivating examples from the introduction. In the case of adoption
of a new technology we can clearly see that a social planner has incentives to intervene.
For example when cities choose a way to make their public transport more eco-friendly
the social planner can have one of the following two interests:
1. He wants to maximize the inter-city transport, i.e. maximize the investment in one
project. In that case electric buses from city 1 can go to city 2 and be charged there. This
avoids pollution and additionally to the eco-friendly intra-city transport, this increases
the benefits for the environment.
2. He wants all cities to invest, i.e. minimize the abstention. More cities will provide
eco-friendly public transport and the pollution decreases.
We give the social planner three different instruments to achieve his goal: He can
change the degree of farsightedness, the taste of a player or modify the threshold of a
project.
Maximize investment in one project A social planner who wants to maximize
the investment in a single projects tries to maximize the following function: U(p) =
maxQ∈P |{i ∈ N : pi = Q}|.
We return to our analysis of Example 6 with the same types of players and thresh-
olds EA = 3 and EB = 2. There are six different stable strategy profiles. For sim-
plicity we rename them: p1 = {A,A,A,A,A,A, ∅}, p2 = {∅, B, B,B,B,B,B}, p3 =
{A,A,A,A,B,A,B}, p4 = {A,A,A,B,A,B,B}, p5 = {A,A,A,B,B,B,B} and p6 =
{∅, ∅, ∅, ∅, ∅, ∅, ∅}. We get U(p1) = U(p2) = 6, U(p3) = 5, U(p4) = U(p5) = 4 and
U(p6) = 0. We denote the profile in which all players select A (B) by pA (pB).
The profiles p1 and p2 are the best stable profiles for the social planner, because all
moderate players coordinate on a project. To move to full coordination, i.e. to pA or
pB the social planner has to turn one of the extremists into a moderate player. From
profile p3 the social planner can improve either by increasing the threshold of project B
to EB = 3 (moving to profile p1) or by turning player 7 into a moderate player (yielding
in profile pA). The difficulty arises in p4. Changing the degree of farsighted or taste of one
player or increasing or decreasing the threshold of one project by 1 does not destabilize
this profile. Either several changes are necessary or for example an increase of EB to 4
(moving to profile p1). Even turning player 7 moderate will not destabilize the profile.
To move from p5 to a preferred profile (p2) the social planner can increase the threshold
of project A to EA = 4. As stated in Proposition 7 the profile p6 is stable if there are too
few farsighted players relative to the thresholds of the project. Either turning any player
except player 1 farsighted or lowering EB to 1 would make the players move to profile p2.
18
We see that different instruments can be used to destabilize certain profiles. Moreover
destabilizing one profile might make the destabilization of another profile more difficult.
In order to have a set of stable strategy profiles that only consist of p1 and p2, the social
planner has to use several instruments. For example, he could increase the thresholds
of both project to EA = EB = 4 and turn three players farsighted. Full coordination
can only be reached when there do not exist two players that are extremists for different
projects.
Finally, notice that making players farsighted does not always benefit the social plan-
ner. When all players are myopic and the planner makes player 6 farsighted, the profile
p1 is destabilized and the players move to profile p4.
Minimizing abstentions If the social planner wants to minimize the amount of players
who abstain, he tries to minimize: V (p) = |{i ∈ N : pi = ∅}|. We use the same example
as before and see that in the profiles p3, p4 and p5 there is no abstention, which is the
most preferred outcome of the planner, i.e. V (p3) = V (p4) = V (p5) = 0. In p1 and p2
one player abstains (V (p1) = V (p2) = 1), while p6 is the worst for the social planner with
V (p6) = 6.
To move from p1 the social planner can either turn player 6 farsighted and so end up
in p4, or turn player 7 into a moderate which yields to pA. On the other hand, from p2
there is no way out besides turning player 1 moderate (or reducing EA to 1). For p6 the
possibilities for this social planner are the same as for the other type of social planner.
This example shows that it is hard to avoid abstention as long as there is a small
amount of extremists for different projects. Reducing the thresholds of all projects to 1 is
a very effective instrument. Then no player has an incentive to abstain. Returning to our
example of cities and eco-friendly transport, this is equivalent to subsidizing all projects.
This might be too expensive for the social planner and in that case other interventions,
as described above, are more suitable.
8 Discussion
We have proposed a model to generalize coordination games on networks. Players want
to coordinate with as many neighbors in their network as possible. In addition, players
not only differ in their taste between the different projects, they also can be either myopic
or farsighted. To analyze this game, we introduced a new stability concept. The set of
stable strategy profiles is equivalent to the set of Nash equilibria as long as all players are
myopic. Introducing farsighted players, i.e. players who anticipate changes that follow
because of their actions, works as a refinement of the set. We have shown that even if all
players are farsighted there exists at least one stable strategy profile. Furthermore, we
have characterized the changes in the set of stable strategy profiles that arise when the
taste of a player is changed.
19
The different projects, that the players can choose between, also add a novelty to the
setting: Each project has a threshold which needs to be reached to make it successful.
Our model allows for ignoring that threshold, i.e. setting all thresholds to 1, but the
majority of our results hold for any level of thresholds. We have shown and illustrated
how the thresholds influence the set of stable strategy profiles.
An important part in our model is the network. Players like to coordinate with their
neighbors in the network. We have analyzed classical network examples such as the
complete network or the star network.
We have discussed the role of a social planner and analyzed two different incentives
for him. A social planner who wants to maximize investment into one project will behave
differently than a social planner who wants to minimize abstention. Changing the type
of a player or the threshold of a project can be good for one type of social planner and
bad for the other one.
Appendix
For the proof of Proposition 1 we require two lemmas:
Lemma A1. Consider a strategy profile p with pi 6= pj for some i 6= j. If at p, player
i prefers p′i to pi (i.e., ui(p′i, p−i) > ui(pi, p−i)) and at (p′i, p−i), player j prefers p′j to
pj (i.e., uj(p′i, p
′j , p−i,j) > ui(p
′i, p−i)) with p′j 6= p′i, then it also holds that at p, player j
prefers p′j to pj (i.e.,uj(p′j, p−j) > uj(pj, p−j)).
Proof. We prove that uj(p′j , p−j) ≥ uj(p
′i, p
′j, p−i,j) > uj(p
′i, p−i) ≥ uj(pj , p−j) holds.
uj(p′j, p−j) ≥ uj(p
′i, p
′j, p−i,j) follows directly from the fact that p′i 6= p′j. If pi is equal to
p′j, then the inequality is strict, otherwise both terms are the same, because player i’s
choices (pi and p′i) both differ from p′j.
It remains to be shown that uj(p′i, p−i) ≥ uj(pj, p−j) holds. We use the same arguments
as before: p′i might be equal to pj , but pi is never equal to pj.
The following lemma follows almost immediately and is presented without proof.
Lemma A2. Consider a strategy profile p from which player i wants to deviate to p′i and
some other player(s) j, i 6= j, wants to deviate to p′j, with p′j 6= p′i; i.e., ui(p′i, p−i) >
ui(pi, p−i) and uj(p′j, p−j) > uj(pj, p−j). Once the deviations have taken place, if player
i deviates again to pi (because ui(pi, p′j, p−i,j) > ui(p
′i, p
′j , p−i,j)), then player j will never
deviate back to pj (since uj(pi, p′j, p−i,j) > uj(pi, pj, p−i,j)).
Proof of Proposition 1. Suppose there exists a myopic-farsighted improving path
p0, p1, ...., pL with p = p0 and p = pL. First, notice that no farsighted player will initiate
such a path. Only myopic players could move away from p. Let i1 be the myopic player
that deviates from p0 to p1, with ui1(p1i1, p0−i1
) > ui1(p0i1, p0−i1
). Consider now the move of
player i2 from p1 to p2. If p2i2 = p1i1, then player i1 will not deviate back to p0i1 since the
20
support for the project she prefers is now even larger. Thus, as long as, along the path,
the successive deviating players continue supporting projet p1i1, player i1 will not deviate
back to p0i1 and then the end strategy profile of the improving path will never be equal to
p.
Let l be the first time along the path such that some myopic player j deviates from pl to
pl+1, with pl+1
j 6= p1i1 = ... = plil and uj(pl+1
j , p1i1, ..., plil, p0−j,i1,...,il
) > uj(p1i1, ..., plil, p
0−i1,...,il
).
Then by Lemma A1, we also have that player j prefers pl+1
j to p0j = p1j = ... = plj before
players i1, ...., il deviates from p0i1 , ..., pl−1
ilto p1i1 = ... = plil 6= pl+1
j . It holds that:
uj(pl+1
j , p1i1, ..., pl−1
ii−1, plil, p
0−j,i1,...,il
) > uj(p1i1, ..., plil, p
0−i1,...,il
),
uj(pl+1
j , p1i1, ..., pl−1
il−1, p0−j,i1,...,il−1
) > uj(p1i1, ..., pl−1
il−1, p0−i1,...,il−1
),
....,
uj(pl+1
j , p0−j) > uj(p0j , p
0−j).
Notice that, in order for the path to end at p, players i1, ...., il−1, il should have in-
centives to deviate back from p1i1 = ... = plil to p0i1 , ..., pl−1
il, respectively. Thus, assume
that this is the case once player j (and possibly other players that deviate later on to the
same project as player j) deviates from pl to pl+1, with pl+1
j 6= p1i1 = ... = plil. But then,
by Lemma A2, we have that player j (and possibly some other player) will never deviate
back to p0j 6= pl+1
j . Hence, p will never be reached.
Suppose now that, at l, a farsighted player j deviates from pl to pl+1, with pl+1
j 6=
p1i1 = ... = plil and uj(pLj , p
L−j) > uj(p
1i1, ..., plil, p
0−i1,...,il
). Then, in order to induce player i1
to deviate back to p0i1 , two cases have to be considered. First, if p0j = p1i1 = ... = plil 6= ∅,
we have that
uj(p0
j , p0
−j) ≤ uj(p1
i1, p0j , p
0
−j,i1) ≤ ... ≤ uj(p
1
i1, ..., plil, p
0
j , p0
−j,i1,...,il) < uj(p
Lj , p
L−j) = uj(p
0
j , p0
−j),
a contradiction.
Second, if p0j = ∅ or p0i1 , ..., p0il6= p0j 6= p1i1 = ... = plil then we have that
uj(p0
j , p0
−j) = uj(p1
i1, p0j , p
0
−j,i1) = ... = uj(p
1
i1, ..., plil, p
0
j , p0
−j,i1,...,il) < uj(p
Lj , p
L−j) = uj(p
0
j , p0
−j),
a contradiction. Thus, no farsighted player along the path from p to pL = p would be the
first player deviating from pl to pl+1, with pl+1
j 6= p1i1 = ... = plil. �
Proof of Proposition 2. We start with the ”if” part of the proof and show that the
set of Nash equilibrium strategy profiles N (g, θ, E) is a set of myopic-farsighted stable
strategy profiles Z(g, θ, E).
Let some p∗ ∈ N (g, θ, E) be given. It holds by definition of the Nash equilibrium that
h(p∗) = ∅. Then, p∗ ∈ N (g, θ, E) is myopic-farsighted stable.
We continue with the “only if” part of the proof and show that N (g, θ, E) is the
only set of myopic-farsighted strategy profiles Z(g, θ, E). Let some p /∈ N (g, θ, E) be
given. Then, by Corollary 1, we have that there exists a Nash equilibrium strategy profile
p∗ ∈ N (g, θ, E) such that p∗ ∈ h(p). Thus, a strategy profile that is no Nash equilibrium
21
cannot be myopic-farsighted stable. �
Proof of Proposition 3. Proof by contradiction:
Assume there is a strategy profile z ∈ Z(g, ((tQi )rQ=1, f), θ−i, E) and that
z /∈ Z(g, ((tQi )rQ=1, m), θ−i, E). This means that there is another profile p = (pi, p−i) with
p−i = z−i such that ui(p) > ui(z). So that the deviation from zi to pi is an immediate
improvement for player i. Then, there is a myopic-farsighted improving path of length 1
from z to p when player i is myopic, with p ∈ h(z).
But then, the previous myopic-farsighted improving path of length 1 from z to p is also
a myopic-farsighted improving path of length 1 from z to p when player i is farsighted.
This implies that z is not stable if i is farsighted and contradicts our assumption. �
Proof of Proposition 4. Let us consider all the strategy profiles p ∈ Z(g, ((tQi =
1/r)rQ=1, si), θ−i, E) with pi = ∅. The only reason why player i chooses to abstain is that
no project is successful. These strategy profiles are still stable when player i becomes an
extremist for any of the projects.
In the remaining profiles of Z(g, ((tQi = 1/r)rQ=1, si), θ−i, E) player i does not abstain.
Let us consider a strategy profile p ∈ Z(g, ((tQi = 1/r)rQ=1, si), θ−i, E) where pi = Q. This
strategy profile can only be stable if project Q is successful because, otherwise, player i
would abstain. This implies p ∈ Z(g, ((tQi = 1, (tRi = 0)R6=Q), si), θ−i, E). Repeating the
same arguments for any other strategy profile where the moderate player i chooses some
of the different projects completes the proof. �
Proof of Proposition 5. Let ℓ be an extremist. Wlog assume that i, j ∈ Nℓ and there
exists no other path between i and j (otherwise repeat the same arguments). Clearly pℓ
is independent of pi and pj , because ℓ is an extremist and always choose the same project.
This implies that pi and pj are independent of each other, that ps ∈ Ni is independent of
pj, that pt ∈ Nj is independent of pi and so on. �
Proof of Proposition 6. We show how an increase of the threshold EQ to EQ +1 can
influence the set of stable strategy profiles. We have to differ between three cases: First,
we characterize the profiles that are stable when the threshold is EQ and that are also
stable for the threshold EQ+1. Second, to compare the two sets of stable profiles for the
thresholds EQ and EQ +1, we have to remove those profiles from Z(g, θ, (EQ, (ER)R6=Q))
that are not stable after increasing the threshold. Third, we have to add new profiles that
are stable under the new threshold EQ + 1 but not stable when the threshold is EQ.
1. and 2. Let us consider p ∈ Z(g, θ, (EQ, (ER)R6=Q)). If we increase the threshold of
project Q by one, there are two possible results. Either the success of all projects (includ-
ing Q) stays unchanged or project Q is no longer successful. In the first case, the profile
p is still stable, because the investment thresholds do not influence the utility, only the
success of the projects do. In the second case, there is not enough investment into project
22
Q. Then for all players who has chosen that project it is better to abstain or invest into
project R 6= Q, so there exists an improving path and p is not stable. This can only
happen if before the increase of the threshold exactly EQ players have chosen project Q.
From this we get Z(g, θ, E) \ {p ∈ Z(g, θ, E) : |{i ∈ N : pi = Q}| = EQ}.
3. We show that a profile that is stable after the increase of the threshold, but was not
stable before the increase is an element of {p : pi = ∅ if tQi = 1 and pi ∈ P\Q otherwise}.
It is impossible that project Q (with threshold EQ + 1) is successful, because than
p ∈ Z(g, θ, (EQ + 1, (ER)R6=Q)) implies p ∈ Z(g, θ, (EQ, (ER)R6=Q)) and p would be stable
under the lower threshold EQ as well. So the only possibility is that project Q is not
successful and that no player will choose project Q. The extremists for project Q will
abstain and all other players will either select a project different than Q or abstain as
well. The set of these profiles is {p : pi = ∅ if tQi = 1 and pi ∈ P\Q otherwise}. �
Proof of Proposition 7. If: Assume for all projects Q ∈ P it holds that |{i ∈ F |tQi >
0}| + min(
1, |{i ∈ M |tQi > 0}|)
< EQ. This implies that for no project Q there are
neither enough farsighted players (with positive taste for project Q) such that they can
either reach the threshold EQ nor are there enough farsighted players to reach the level
EQ − 1 and one myopic player that has positive taste for project Q. So the farsighted
players will not start an improving path from p = {∅, ∅, ..., ∅}. Clearly, the myopic players
will not deviate either. This implies h({∅, ∅, ..., ∅}) = ∅ and p = {∅, ∅, ..., ∅} is stable.
Only if: Proof by contradiction: Assume that there is a project R ∈ P with |{i ∈ F |tRi >
0}| + min(
1, |{i ∈ M |tRi > 0}|)
≥ ER. There exists the following improving path from
p = {∅, ∅, ..., ∅} to p′ with p′i = R for all players i with tRi > 0. First all farsighted players
(with positive taste for project R) will deviate from pi = ∅ to p′i = R. By assumption
there are at least ER − 1 such players. Once ER − 1 farsighted players have selected
project R, another player with tRi > 0 will deviate to project R as well and make project
R successful. This can either be a myopic player or a farsighted player. The existance of
this improving path implies that p = {∅, ∅, ..., ∅} cannot be in this case. �
Proof of Lemma 1. Follows from Lemma A1 and the fact that the utility of player i
who picks project pi is the same when her neighbor invests into project pj 6= pi or p′j 6= pi.
�
Proof of Proposition 8. We show by contradiction that there cannot be some mod-
erate players that choose project Q and some moderate players that choose project R.
The proof is easily replicable to show the same result for three or more projects. Let us
assume there exist moderate players r1, . . . , rz who choose project R and players q1, . . . , qy
who choose project Q 6= R.
It is obvious that players deviate if the threshold of either project Q or R is not reached.
For the remainder we assume that both projects are successful.
Wlog, assume that Q = |XQ∪{q1, . . . , qz}| ≥ |XR∪{r1, . . . , ry}| = R. Notice that, in the
23
complete network, for all i ∈ {r1, . . . , ry} : |{j ∈ Ni(g) | pj = Q}| > |{j ∈ Ni(g) | pj = R}|.
This implies that ui(pi = Q) > ui(pi = R). Since Q > R−1 it follows that a myopic player
i ∈ {r1, . . . , ry} will deviate to project Q. Farsighted players anticipate that if they choose
Q, the moderate players will follow them and so they know that Q = N−1−|XR| > R−1
holds and they will change to project Q as well. �
Proof of Proposition 9. If: Let us assume that player i ∈ {2, . . . , n − 1} is myopic
and players j 6= i ∈ {2, . . . , n−1} are farsighted. Obviously the strategy profiles in which
all players choose the same action are stable. We show that no other strategy profile can
be stable:
Case q: Let us assume that two farsighted neighbors j, k (with k = j+1) choose different
projects and all ℓ < j choose project pj and all ℓ > k choose pk. Wlog we can assume that
i > k, i.e. pi = pk. Then player j has an incentive to change her decision and to select
project pk. If player j changes her project, then player ℓ = j − 1 changes her project,
because she knows that her neighbor m = j−2 will change. Player j−x changes, because
the farsighted player 2 who is the neighbor of the loose end knows that if she changes her
decision the loose end will follow.
Case 2: Let us (wlog) assume that player i chooses project Q and all players located to
her left do the same, i.e. ∀j ∈ N with j < i : pj = pi = Q while all other players choose
R. In this case the neighbor of player i who chooses a different project than i, namely
player i+ 1 (with pi+1 = R) has an incentive to deviate. Player i+ 1 changes because all
other players i + 2 to n − 1 will change. These players change, because they anticipate
that player n (the loose-end) will change as well.
So under these assumptions there can be no other stable strategy profile.
Only if (for EQ = 1 for all Q ∈ P ): We assume there are two myopic players i, j in
{2, . . . , n− 1}.
1. Case: Assume i and j are neighbors, i.e. j = i+1. Then there exists a stable strategy
profile in which all players k ≤ i choose project Q and all players ℓ ≥ j = i + 1 choose
project R 6= Q. In that case the farsighted players 2 to i− 1 and player 1 (whose degree
of farsightedness does not matter) chose Q and the farsighted players j + 1 to n− 1 and
player n chose project R. This implies that neither of the players 1 to i− 1 and j + 1 to
n has any incentive to deviate. The two myopic players have also no incentive to change,
because their utility after changing would be the same as before.
Case 2: Assume i and j are not neighbors, wlog we assume i < j.
We show that the following strategy profile is stable: ∀k < j : pk = Q and ∀k ≥ j : pk = R.
In this strategy profile there are some farsighted players between i and j who choose
project Q. The myopic player j has no incentive to deviate to project Q, because one
of her neighbors chooses Q and the other chooses R. The farsighted player j − 1 would
choose project R only if j − 2 chooses project R as well. Player j − 2 would deviate
to project R only if j − 3 deviates to R as well, continuing this arguments leads to the
problem that i + 1 only chooses project R if i deviates to R. Since i is myopic she does
24
not deviate, because even if i+1 choose R, she still get the same utility from the projects
Q and R. So the strategy profile described is stable. �
References
Allouch, N. (2015) On the private provision of public goods on networks Journal of Eco-
nomic Theory 157:527-552
Ballester, C. and Calvo-Armengol, A and Zenou, Y. (2006) Who’s who in networks.
Wanted: The key player Econometrica 74.5:1403-1417
Bramoulle, Y. (2007) Anti-coordination and social interactions Games and Economic
Behavior 58.1:30-49
Bramoulle, Y. and Lopez-Pintado, D. and Goyal, S. and Vega-Redondo, F. (2004) Network
formation and anti-coordination games International Journal of Game Theory 33.1:1-19
Bramoulle, Y. and Kranton, R. (2007) Public goods in networks Journal of Economic
Theory 135.1:478-494
Chwe, M. (1994) Farsighted Coalitional Stability Journal of Economic Theory 63.2:299-
325
Fudenberg, D. and Tirole, J. (1991) Game Theory. MIT Press
Goyal, S. and Vega-Redondo, F. (2005) Network formation and social coordination Games
and Economic Behavior 50.2:178-207
Harsanyi, J. (1974) An Equilibrium-Point Interpretation of Stable Sets and a Proposed
Alternative Definition Management Science 20.11:1472-1495
Herings, P.J.J. and Mauleon, A. and Vannetelbosch, V. (2009) Farsightedly stable net-
works Games and Economic Behavior 67.2:526-541
Jackson, M. O. and Storms, E. (2018) Behavioral communities and the atomic structure
of networks
Jackson, M. O. and Watts, A. (2002) On the formation of interaction networks in social
coordination games Games and Economic Behavior 41.2:265-291
Jackson, M. O. and Watts, A. (2002) The evolution of social and economic networks
Journal of Economic Theory 106.2:265-295
Leister, C. M. and Zenou, Y. and Zhou, J. (2018) Coordination on networks (May 29,
2018)
25
Luce, R. D. and Raiffa, H. (1957) Games and decisions: Introduction and critical survey.
Oxford, England: Wiley
Ray, D. and Vohra, R. (2015) The farsighted stable set Econometrica 83: 977-1011
Sakovics, J. and Steiner, J. (2012) Who matters in coordination problems? The American
Economic Review 102:3439-3461
26